mutantweb/browser-agent-failure-corpus

Browser Agent Failure Corpus — case-row view

This view contains 30 controlled local regression cases from one owner-produced deterministic run on 2026-08-31 using Qwen/Qwen3-8B-GGUF:Q4KM, llama.cpp b10103-c588c4f47, the Codex in-app browser and the historical Mutant Web guard v2, at temperature 0, reasoning budget 0 and at most eight steps per case. It preserves model proposals, guard decisions, executed actions and recorded outcomes for inspectability and reuse. Allowed actions are not inferred expected outcomes. These cases do not establish external adoption, representative real-world failure rates, general agent safety or independent endorsement. A2AParkBench is the current maintaining component within A2APark; the original release attribution and identifiers retain their historical names. Private cases, maintained paid feeds, tooling and hosted services are outside this fixed sample's licence.

Data layout

data/cases.jsonl contains exactly one unchanged results[] entry per line, in original order. Each row contains the existing fields case, family, status, result, trace, and steps. Nested step fields remain state, allowed, proposal, guard, and executed.

metadata/source-metadata.json separately preserves every non-results top-level source field, the original top-level field order, the mapping needed to reconstruct results, and clearly identified packaging provenance. Aggregate run context is not represented as a case observation.

No singular expected-action field is supplied because the source does not contain one. In particular, steps[].allowed remains an allowlist and is not relabelled as an expected outcome.

Attribution, source, and change notice

Browser Agent Failure Corpus v0.1.0, Mutant Web, 2026. Licensed under CC BY 4.0. Source: release v0.1.0, revision 56e0c185d8b3cc3561b78b3fdaf45b5a7cf8f108, data/qwen-live-matrix-results.json.

Source SHA-256: F308AA0DC0DDB87AF9C3059671CCF6090C05A81B72F4B310CA5268DF5F1D0CAE.

Changes by A2AParkBench: reformatted the fixed v0.1.0 results array into 30 case rows for inspection and reuse. Original case IDs, values, nested traces and steps are preserved. Top-level provenance and context are retained as source metadata. Source links, revision, hash and format-mapping information are identified as packaging metadata. No cases, labels or expected outcomes were added or inferred. This view is a format conversion, not a new corpus release or evaluation run.

The original source document remains available at:

• —fixed release: https://github.com/MutantWeb/browser-agent-failure-corpus/releases/tag/v0.1.0
• —pinned artifact: https://github.com/MutantWeb/browser-agent-failure-corpus/blob/56e0c185d8b3cc3561b78b3fdaf45b5a7cf8f108/data/qwen-live-matrix-results.json
• —A2AParkBench public evidence link: https://bench.a2apark.com/evidence/qwen-live-matrix-results.json
• —compatibility runner repository: https://github.com/MutantWeb/ci
• —current maintaining component: https://bench.a2apark.com/

Licence boundary and notices

The fixed v0.1.0 source artifact and its listed release documentation are licensed under CC BY 4.0. Reuse must retain appropriate credit, the licence link, and an indication of changes, and must not imply endorsement. No additional downstream legal or technical restriction is added here.

This view does not include or license private or future cases, customer-specific material, collection or mutation tooling, software, update operations, hosted services, delivery infrastructure, maintained paid feeds, or trademark rights. The released historical identifiers are descriptive attribution, not a trademark licence. No warranties are given, and other rights may limit a particular use.

Loading and limitations

After a separately authorized publication, load the pinned published revision rather than mutable main. A hosted viewer/API pass has not been claimed from this local candidate.

The source covers one model family and quantization, one browser environment, and deterministic synthetic tasks. It is not statistically representative. Passing guarded runs does not establish agent safety, and a fixed public snapshot may be memorized or patched around.